betaFeedback

WorksComputer Science

TACO: Trajectory-Aware Curvature Observation for Memorization Localization and Mitigation in Diffusion Models

Abstract

Text-to-image diffusion models can reproduce memorized training content within localized image regions, making precise localization essential for targeted mitigation. However, curvature observations at a single denoising state can contain transient responses, while spatially uniform mitigation may disrupt prompt-relevant content. We present Trajectory-Aware Curvature Observation (TACO), a unified framework for local memorization detection and mitigation in diffusion models. Our empirical analysis reveals that memorization-related curvature becomes pronounced during late-stage denoising and persists at consistent spatial locations across neighboring states. Motivated by these observations, TACO aggregates prompt-dependent curvature responses along the late denoising trajectory and applies latent-guided spatial refinement to improve spatial coherence. The resulting memorization map serves both as a localization signal and as a spatial control signal for mitigation. Specifically, TACO uses this map to steer existing inference-time interventions toward memorization-related regions through spatial reweighting and localized guidance correction. Extensive experiments demonstrate that TACO consistently outperforms competing localization methods and improves the Pareto trade-off between memorization suppression and generation quality. These results establish trajectory-aware spatial analysis as an effective link between memorization localization and targeted mitigation.

Introduction

Text-to-image (T2I) diffusion models [12, 29, 30] can reproduce training images or distinctive visual patterns under specific prompts [3, 26, 32]. Such memorization raises privacy and copyright concerns, as generated outputs may expose sensitive training content or reproduce protected material.

Such memorization can occur globally or locally [5, 17]. Global memorization closely reproduces a training image as a whole, while local memorization retains copied content within part of an otherwise varying image. Figure 1 illustrates both cases: the global example exhibits whole-image copying, whereas the local example preserves memorized content despite changes in wall color. These variations can make local copying less apparent, highlighting the need to identify its spatial extent. Localization also enables selective mitigation: interventions can target copied regions while preserving surrounding content, reducing the unnecessary changes that image-wide suppression may introduce.

Global and local memorization examples

Figure 1. Global and local memorization in Stable Diffusion v1.4 [24] across 6 different seeds.

Existing detection methods provide useful evidence of copying, but only limited efforts have been devoted to localizing where memorized content occurs within generated images. Retrieval-based methods compare generated images against training examples [3, 21, 26]; retrieval-free methods analyze model responses, such as conditional-unconditional score differences [14, 25, 33]. These approaches can indicate whether an output is memorized, but a single image-level score does not show which region is responsible. Attention-based methods use cross-attention to identify potentially copied regions [5, 23], and coordinate-wise curvature differences offer a complementary geometric signal for localization [15, 16]. However, a map estimated at one denoising state may include transient responses and background noise. A reliable local detector should therefore capture evidence that persists along the denoising trajectory.

This requirement is closely tied to inference-time mitigation. Prior methods modify prompts or embeddings [27, 33], adjust initial noise [9], or change classifier-free guidance (CFG) during sampling [11, 13]. Dynamic attention intervention offers another way to selectively alter text conditioning [35]. Although these methods intervene at different stages of generation, applying their changes globally can alter non-memorized regions alongside copied content. Therefore, a spatial memorization signal can help direct existing interventions toward locally copied regions, suppressing memorization while preserving the surrounding content.

Motivated by these, we present Trajectory-Aware Curvature Observation (TACO), a unified framework for local memorization detection and mitigation. Our empirical study shows that memorization-related curvature becomes pronounced in late denoising and remains spatially consistent across neighboring states. TACO uses this property to estimate prompt-dependent coordinate-wise curvature at multiple late-stage states, refine each response using the latent’s spatial structure, and aggregate the observations into a continuous memorization map. The map serves two roles: thresholding it yields a localization mask, while retaining its continuous values provides spatial weights for mitigation. In mitigation, TACO separates each intervention into a component aligned with a reference direction, which preserves the baseline objective, and a component that can be spatially adjusted. When the baseline objective gradient is available, TACO uses it as the reference direction, preserving the aligned component and spatially redistributing the orthogonal component according to the memorization map. When it is unavailable, TACO uses the score difference as the reference, combining map-weighted CFG attenuation with a transverse correction orthogonal to that direction.

We evaluate TACO on Stable Diffusion (SD) models [24] using benchmarks for local-region detection and memorization mitigation. TACO improves localization over competitive baselines, and its integration with diverse existing mitigation methods advances the Pareto frontier [18], achieving a more favorable trade-off between memorization suppression and image quality, as measured by text alignment and perceptual quality. Our contributions are threefold:

  • We introduce TACO, a unified framework for memorization localization and mitigation in T2I diffusion models that advances the Pareto frontiers between memorization suppression and generation quality.

  • We leverage the temporal persistence of late-stage curvature, combining coordinate-wise curvature estimation, latent-guided refinement, and trajectory aggregation to produce spatially coherent memorization maps.

  • We develop map-guided steering for existing inference-time interventions and validate TACO’s localization and mitigation benefits across multiple baselines and Stable Diffusion model variants.

Method

Preliminaries

We consider a latent-space text-to-image diffusion model that generates an image through an iterative denoising process with TT sampling steps. Throughout this paper, step refers exclusively to an actual denoising update during inference, whereas timestep refers to the diffusion-time coordinate used by the scheduler.

Let x0∈RC×H×Wx_0 \in\mathbb{R}^{C \times H \times W} denote the initial Gaussian noise, and let xt∈RC×H×Wx_t \in\mathbb{R}^{C \times H \times W} denote the latent state obtained after tt steps. The resulting sampling trajectory is therefore {xt}t=0T\{x_t\}_{t=0}^{T}, where xTx_T is the final denoised latent that is subsequently decoded into the generated image.

Given a text prompt pp, at each sampling step tt, the denoising network produces an unconditional noise prediction ϵtu(xt)\epsilon_t^u(x_t) and a conditional prediction ϵtc(xt,p)\epsilon_t^c(x_t,p). Classifier-free guidance combines them as

ϵtg(xt,p)=ϵtu(xt)+s(ϵtc(xt,p)−ϵtu(xt)),(1)\epsilon_t^g(x_t,p) = \epsilon_t^u(x_t) + s\left(\epsilon_t^c(x_t,p) - \epsilon_t^u(x_t)\right), \tag*{(1)}

where ss is the guidance scale. The scheduler then applies the corresponding denoising transition to obtain the next latent state. Thus, increasing tt corresponds to progressing along the inference-time denoising trajectory from x0x_0 toward the final generated latent xTx_T.

Our study considers two closely related tasks: memorization localization and memorization mitigation. The localization task identifies regions of the generated image that contain memorized content using signals from the denoising trajectory. The mitigation task uses these regions to guide inference-time interventions, reducing the reproduction of memorized training content while preserving image quality and semantic alignment. For both tasks, TACO produces a continuous spatial memorization map M∈RH×WM \in\mathbb{R}^{H \times W}, where a larger value at spatial location xx indicates stronger evidence of memorization-related activity. For localization, MM is converted into a binary mask to determine the spatial extent of memorized content. For mitigation, the same continuous map is retained as a spatial control signal to guide the intervention along the denoising trajectory.

TACO for Memorization Localization

The key idea of TACO is to identify memorization from curvature patterns that persist along the sampling trajectory. Here, curvature measures the coordinate-wise second-order sensitivity of the denoising residual to the latent state. Given a prompt pp, TACO observes curvature at multiple states {xt}t∈Tobs\{x_t\}_{t\in\mathcal{T}_{\mathrm{obs}}} along the trajectory. By preserving spatial correspondence and aggregating responses across these states, TACO captures persistent curvature evidence while reducing the influence of transient responses.

Empirical Evidences for TACO. We first investigate how curvature responses evolve along the sampling trajectory and whether their spatial patterns remain stable across neighboring states. All analyses use a sampling process with T=49T = 49, and are conducted over 200 samples. For each step, we summarize the curvature map by averaging its strongest 5% spatial responses, which gives a trajectory profile of peak curvature intensity. As shown in Figure 2(a), the curvature response of memorized samples increases sharply toward the end of sampling. In comparison, the response of non-memorized samples increases only mildly and remains substantially lower. This observation indicates that the curvature associated with memorization becomes particularly pronounced during the late stage of sampling, making these states especially informative for localization.

Figure 2. Empirical studies motivating trajectory-aware curvature observation. (a) Evolution of peak curvature intensity along the sampling trajectory. (b) Spatial consistency of curvature responses between neighboring late-stage states. (c) Signal-to-background ratio and background variance of individual and fused late-stage curvature responses.

We next examine whether the spatial locations of strong curvature responses remain stable across neighboring late-stage states. Specifically, we compute the spatial cosine similarity between the curvature maps at t=48t = 48 and t=49t = 49. As shown in Figure 2(b), the strong responses associated with memorization tend to persist at consistent spatial locations across neighboring states, providing a reliable temporal correspondence for trajectory-level aggregation.

Finally, we investigate whether combining these neighboring observations improves the stability of the resulting spatial signal. We average the curvature maps at t=48t = 48 and t=49t = 49 and measure their signal-to-background ratio (SBR), where the strongest 10% of responses define the signal region and the remaining locations form the background. We additionally measure the background variance to quantify fluctuations outside the signal region. In Figure 2(c), the fused response maintains a high SBR while substantially reducing the background variance compared with the highly responsive yet noisy observation at t=49t = 49, providing a favorable balance between signal contrast and background stability.

Together, these findings show that memorization-related curvature strengthens and becomes spatially consistent in late-stage sampling, while aggregation across neighboring states improves background stability. This motivates TACO to model curvature as a trajectory-level spatial signal.

Trajectory-Aware Curvature Signal. Building on coordinate-wise curvature differences [16], we first extract the prompt-dependent curvature response at each observation state. For a given state xtx_t, we isolate the prompt-dependent component of the denoising dynamics as

rt(xt,p)=ϵtc(xt,p)−ϵtd(xt).(2)r_t(x_t,p) = \epsilon_t^c(x_t,p) - \epsilon_t^d(x_t). \tag*{(2)}

We then estimate the diagonal curvature of the normalized residual rt/σtr_t/\sigma_t, where σt=1−αˉt\sigma_t = \sqrt{1-\bar{\alpha}_t} is the noise scale at state tt. We use JJ structured Walsh-Hadamard probes {v(j)}j=1J\{v^{(j)}\}_{j=1}^{J} [2], with fixed spatial assignments, channel-dependent offsets, and a shared random sign field. These probes provide approximately orthogonal coverage of the latent coordinates. The resulting curvature estimator is

H^t=1J∑j=1Jv(j)⊙∇xt⟨rt(xt,p)σt,v(j)⟩.(3)\widehat{H}_t = \frac{1}{J}\sum_{j=1}^{J} v^{(j)} \odot\nabla_{x_t}\left\langle\frac{r_t(x_t,p)}{\sigma_t}, v^{(j)} \right\rangle. \tag*{(3)}

where H^t(x,c)\widehat{H}_t(x,c) denotes the curvature at spatial location xx and channel cc. We retain the negative response Ht−(x,c)=ReLU⁡(−H^t(x,c))H_t^{-}(x,c) = \operatorname{ReLU}(-\widehat{H}_t(x,c)), and average it across channels to obtain the spatial response Ct(x)=1C∑c=1CHt−(x,c)C_t(x) = \frac{1}{C}\sum_{c=1}^{C} H_t^{-}(x,c).

To emphasize salient curvature responses, we standardize CtC_t across spatial locations and apply sigmoid-based reweighting as Ct(x)=Ct(x)sigmoid⁡(βCt(x)−μtσt+ϵ)C_t(x) = C_t(x)\operatorname{sigmoid}\left(\frac{\beta C_t(x)-\mu_t}{\sigma_t+\epsilon}\right), where μt\mu_t and σt\sigma_t denote the spatial mean and standard deviation of CtC_t, respectively, and β=2.0\beta= 2.0 controls the strength of the reweighting. This operation enhances relatively strong responses while retaining their original magnitudes.

The key difference from single-state curvature localization is that TACO preserves the spatial correspondence of CtC_t across the sampling trajectory. Specifically, for the observation states Tobs\mathcal{T}_{\mathrm{obs}}, we obtain the sequence {Ct}t∈Tobs\{C_t\}_{t\in\mathcal{T}_{\mathrm{obs}}} and retain each response at its original spatial location. This allows curvature evidence to be evaluated jointly across neighboring states, following the trajectory-level persistence observed in our empirical analysis.

Latent-Guided Spatial Refinement. Although observations across neighboring states provide temporal evidence, each state-wise curvature map may still contain isolated spatial peaks. We therefore refine each response CtC_t using the spatial structure of its corresponding latent state xtx_t before trajectory-level aggregation.

Let ft(x)∈RCf_t(x) \in\mathbb{R}^{C} denote the latent feature at spatial location xx. We perform Principal Component Analysis (PCA) over all spatial locations and project the centered features onto the first principal direction ut\mathbf{u}_t as Gt(x)=utT(ft(x)−μt)G_t(x) = \mathbf{u}_t^{\mathsf{T}}(f_t(x)-\mu_t), where μt\mu_t is the spatial mean of the latent features. The resulting map GtG_t summarizes the dominant spatial variation in xtx_t and serves as a lightweight structural cue. For each y∈N(x)y \in\mathcal{N}(x), we define the edge-aware affinity wt(x,y)=exp⁡(−∥x−y∥222σs2)exp⁡(−∣Gt(x)−Gt(y)∣22σg2(x))w_t(x,y)=\exp\left(-\frac{\lVert x-y\rVert_2^2}{2\sigma_s^2}\right)\exp\left(-\frac{\lvert G_t(x)-G_t(y)\rvert^2}{2\sigma_g^2(x)}\right), where σs\sigma_s controls the spatial extent of the neighborhood and σg(x)\sigma_g(x) controls sensitivity to local variations in the guidance map. We then refine the curvature response by normalized local aggregation:

C~t(x)=∑y∈N(x)wt(x,y)Ct(y)∑y∈N(x)wt(x,y).(4)\widetilde{C}_t(x)=\frac{\sum_{y\in\mathcal{N}(x)}w_t(x,y)C_t(y)}{\sum_{y\in\mathcal{N}(x)}w_t(x,y)}. \tag*{(4)}

This refinement suppresses isolated peaks that lack local support, preserves spatially consistent responses, and limits smoothing across latent boundaries. The refined map C~t\widetilde{C}_t is subsequently used for trajectory-level aggregation.

Trajectory-Level Memorization Map. Figure 2(a) shows that memorization-related curvature becomes stronger during the late denoising stages. Based on this observation, we select a short sequence of neighboring late-stage states for curvature observation, which concentrates computation on the stages with the strongest evidence while avoiding initial states that provide limited additional information.

Let Tobs\mathcal{T}_{\mathrm{obs}} denote the selected observation states, with tstartt_{\mathrm{start}} and tendt_{\mathrm{end}} denoting the earliest and latest states, respectively. We aggregate their refined curvature responses using a normalized weighted sum:

M(x)=∑t∈TobswtC~t(x)∑t∈Tobswt,wt=2−t−tstarttend−tstart.(5)M(x)=\frac{\sum_{t\in\mathcal{T}_{\mathrm{obs}}}w_t\widetilde{C}_t(x)}{\sum_{t\in\mathcal{T}_{\mathrm{obs}}}w_t},\quad w_t=2-\frac{t-t_{\mathrm{start}}}{t_{\mathrm{end}}-t_{\mathrm{start}}}. \tag*{(5)}

This weighting scheme assigns greater importance to earlier observations within the selected interval. The motivation is that neighboring states exhibit complementary characteristics as in Figure 2(c): earlier observations can provide more stable spatial responses with lower background noises, whereas later observations can exhibit stronger curvature contrast alongside increased background variation. Trajectory aggregation yields a continuous spatial map MM that summarizes temporally persistent and spatially coherent curvature evidence across the late-stage states. TACO uses this map in two ways: thresholding MM provides a binary localization mask, while its continuous values serve as spatial weights for inference-time mitigation.

TACO-Steered Memorization Mitigation

The continuous TACO map MM provides a spatial signal for steering an existing inference-time memorization mitigation method. At each applicable state, TACO takes the intervention generated by the baseline method and spatially modulates its effect toward regions with stronger memorization responses, while preserving the baseline method’s original mitigation objective.

Let Δtbase\Delta_t^{\mathrm{base}} denote the intervention produced by the baseline method at sampling state tt, and let gtg_t denote the gradient associated with its mitigation objective. We decompose the baseline intervention into an objective-aligned component and an orthogonal tangent component:

Δtbase=Δt∥+Δt⊥,(6)\Delta_t^{\mathrm{base}}=\Delta_t^{\parallel}+\Delta_t^{\perp}, \tag*{(6)}

where Δt∥=⟨Δtbase,gt⟩∥gt∥22gt\Delta_t^{\parallel}=\frac{\langle\Delta_t^{\mathrm{base}},g_t\rangle}{\lVert g_t\rVert_2^2}g_t, and Δt⊥=Δtbase−Δt∥\Delta_t^{\perp}=\Delta_t^{\mathrm{base}}-\Delta_t^{\parallel}. Δt∥\Delta_t^{\parallel} captures the part of the intervention directly aligned with the original mitigation objective, while Δt⊥\Delta_t^{\perp} provides the degrees of freedom that can be spatially redistributed.

TACO applies the memorization map MM only to this redistributable component Δt⊥\Delta_t^{\perp}. At step tt, MM is first resized to the spatial resolution of xtx_t and locally expanded to tolerate small localization errors: Mtdil=MaxPool⁡(Resize⁡(M))M_t^{\mathrm{dil}}=\operatorname{MaxPool}(\operatorname{Resize}(M)). We then normalize MtdilM_t^{\mathrm{dil}} to M~t(x)=Mtdil(x)−Mtmin⁡Mtmax⁡−Mtmin⁡+ϵ\widetilde{M}_t(x)=\frac{M_t^{\mathrm{dil}}(x)-M_t^{\min}}{M_t^{\max}-M_t^{\min}+\epsilon}. The resulting weight assigns stronger intervention to locations with stronger memorization evidence while maintaining a continuous transition across neighboring regions.

Afterwards, Δt⊥\Delta_t^{\perp} is spatially reweighted according to M~t(x)\widetilde{M}_t(x):

Δ^t⊥=M~t⊙Δt⊥.(7)\widehat{\Delta}_t^{\perp}=\widetilde{M}_t\odot\Delta_t^{\perp}. \tag*{(7)}

Since spatial reweighting can change the magnitude of the intervention, TACO rescales Δ^t⊥\widehat{\Delta}_t^{\perp} to preserve the magnitude of the original tangent update as:

Δt⊥,TACO=∥Δt⊥∥2∥Δ^t⊥∥2+ϵΔ^t⊥.(8)\Delta_t^{\perp,\mathrm{TACO}}=\frac{\lVert\Delta_t^{\perp}\rVert_2}{\lVert\widehat{\Delta}_t^{\perp}\rVert_2+\epsilon}\widehat{\Delta}_t^{\perp}. \tag*{(8)}

The final TACO-steered intervention is therefore ΔtTACO=Δt∥+Δt⊥,TACO\Delta_t^{\mathrm{TACO}}=\Delta_t^{\parallel}+\Delta_t^{\perp,\mathrm{TACO}}. This formulation preserves the component required by the original mitigation objective and the overall magnitude of the redistributable intervention, while concentrating its spatial effect according to the TACO memorization map. Consequently, TACO changes where the mitigation acts rather than simply adjusting its global strength.

More generally, the steering operation is independent of the specific form of the underlying mitigation procedure. When the objective gradient gtg_t is unavailable, TACO uses the current score difference rt=ϵtc−ϵtur_t = \epsilon_t^c - \epsilon_t^u as the reference direction for gradient decomposition. Specifically, TACO computes the gradient of a map-weighted attraction objective and projects it onto the subspace orthogonal to rtr_t, yielding a transverse correction that is normalized and scaled before being added to the baseline intervention. In parallel, the TACO map locally attenuates classifier-free guidance in regions with stronger memorization responses. These two components jointly steer the mitigation toward localized memorized content while preserving the orthogonality of the gradient correction to the current score-difference direction, without requiring an explicit gradient from the underlying mitigation objective.

Overall, TACO establishes a direct connection between localization and mitigation: the same continuous trajectory-aware map that identifies where memorization is expressed is reused to determine where the mitigation intervention should be concentrated. Detailed algorithms for localization and mitigation are provided in the supplementary material.

Experiments

Implementation Details

We evaluate memorization localization on SD v1.4 and v2.1 [24], and memorization mitigation on SD v1.4 and v2.0, following the configurations of the corresponding benchmarks. We use DDIM sampling [28] with 50 inference steps and a classifier-free guidance scale of 7.5. TACO observes the last two denoising states, Tobs={48,49}\mathcal{T}_{\mathrm{obs}} = \{48, 49\}, and estimates coordinate-wise curvature using 16 structured Walsh-Hadamard probes. All experiments are conducted on one NVIDIA GeForce RTX 4090 GPU under Ubuntu 22.04.

Benchmarks and Baselines

Memorization Localization. We evaluate localization on the benchmark provide by [16] under two settings: TV only, containing samples with localized memorization, and All, which additionally includes non-memorized samples. Using ground-truth binary masks, we report mean Intersection over Union (mIoU), which measures spatial overlap with memorized regions, and pixel-level accuracy (Acc), which measures the proportion of correctly classified pixels. Higher values indicate better localization.

Baselines include Bright Ending (BE) [5] and four variants from [16]: curvature-difference estimators and their score-difference approximations, using either the unconditional model (cov, score_diff) or a less-trained model (cov_bad, score_diff_bad).

Memorization Mitigation. We evaluate memorization mitigation on the benchmark provided by [32], which covers both local and global memorization. We use SSCD [21] to measure the visual similarity between each mitigated image and its corresponding reference, where lower similarity indicates stronger memorization suppression. To assess whether this suppression preserves generation quality, we additionally use CLIP [22] to measure image-text alignment and ImageReward [34] to predict prompt-conditioned human preference. Higher CLIP and ImageReward scores indicate better semantic alignment and perceptual quality, respectively. We integrate TACO into four baselines: batch-wise and sample-wise initial-noise adjustment [9] (Init Batch, Init Sample), and early-stage CFG perturbations with zero or opposite guidance [13] (Basin Zero, Basin Neg). We vary each baseline’s mitigation strength and compare its Pareto frontiers with and without TACO in the SSCD-CLIP and SSCD-ImageReward spaces to assess the memorization-quality trade-off. Unless otherwise specified, we follow the default configurations of each baseline.

Experimental Results

Memorization Localization. Table 1 reports quantitative localization results on SD v1.4 and SD v2.1 under the TV only and All settings. TACO consistently achieves the best overall performance across both models and evaluation settings, demonstrating its ability to localize memorized regions accurately while distinguishing them from non-memorized content. On SD v1.4, TACO achieves an mIoU of 0.942 and an accuracy of 0.964 on TV only, improving over the strongest competing results by 2.0 and 1.4 percentage points, respectively. On All, it further outperforms the best competing results by 1.9 and 0.9 percentage points. The improvements remain consistent on SD v2.1.

ModelMethodTV only mIoU(↑)TV only ACC(↑)All mIoU(↑)All ACC(↑)
All-zeros0.0000.4370.3180.614
All-ones0.5630.5630.3860.386
SD v1.4BE0.6930.7660.4670.780
cov0.9070.9410.9070.937
cov_bad0.9220.9500.8470.940
score_diff0.8260.8910.8470.906
score_diff_bad0.8570.9110.6450.914
Ours0.9420.9640.9260.949
All-zeros0.0000.3520.2390.507
All-ones0.6480.6480.4930.493
SD v2.1BE0.7940.8520.8150.887
cov0.9380.9580.9080.968
cov_bad0.9420.9630.9020.971
score_diff0.7810.8360.8140.876
score_diff_bad0.9150.9440.8990.957
Ours0.9570.9720.9230.977

Table 1. Quantitative evaluation of memorization localization. Bold and underlined values denote the best and second-best results, respectively.

TACO also maintains a competitive computational cost. Averaged over individual samples, cov, cov_bad, score_diff, and score_diff_bad require 1.87, 2.40, 1.32, and 2.57 seconds, respectively, while TACO requires 2.41 seconds. Although TACO aggregates curvature evidence from two late-stage states, its runtime is nearly identical to that of cov_bad, while achieving markedly better localization performance.

Figure 3 provides a qualitative comparison between TACO and the four baselines evaluated in [16]. TACO produces more spatially coherent responses with sharper boundaries, better capturing memorized regions while suppressing diffuse responses in surrounding non-memorized areas. To further examine the utility of these maps for mitigation, we apply our proposed steering pipeline to Init Batch using each method’s predicted map. The resulting images demonstrate that the more precise localization provided by TACO enables more targeted intervention while better preserving image quality. In contrast, less precise baseline maps may lead to suboptimal mitigation results. This comparison indicates that improved localization not only enhances interpretability but also provides more effective spatial guidance for memorization mitigation.

Qualitative comparison of memorization localization and map-guided mitigation

Figure 3. Qualitative comparison of memorization localization and map-guided mitigation for prompts: “Foyer painted in HABANERO” (top) and “Plymouth Curtain Panel featuring Madelyn - White Botanical Floral Large Scale by heatherdutton” (bottom). For each method, the left panel shows its original localization map, with blue and red indicating weaker and stronger memorization responses, respectively. The right panel shows the mitigated image obtained by applying our proposed steering method with the corresponding localization map.

Memorization Mitigation. We next evaluate whether TACO can improve existing inference-time mitigation methods without sacrificing semantic alignment or perceptual quality. Figure 4 and Figure 5 compare the Pareto frontiers of the original baselines and their TACO-steered counterparts on SD v1.4 and SD v2.0, respectively. Each point corresponds to a different mitigation strength. Across both diffusion models, TACO steering consistently improves the trade-off between memorization suppression and generation quality. In the SSCD-CLIP space, the TACO-steered frontiers extend toward lower SSCD and higher semantic alignment; similar improvements are observed in the SSCD-ImageReward space. Importantly, these gains are observed

Figure 4. SSCD-CLIP and SSCD-ImageReward Pareto fronts in SD v1.4.

Figure 5. SSCD-CLIP and SSCD-ImageReward Pareto fronts in SD v2.0.

across four distinct intervention strategies, including batch-wise and sample-wise initial-noise adjustment and early-stage CFG perturbations with zero or opposite guidance. This consistency indicates that TACO is not tied to a particular mitigation objective, but can spatially redirect existing interventions toward memorization-related regions. The quantitative Pareto-front statistics in the supplementary material further support this observation.

Figure 6 further illustrates the qualitative benefits of TACO steering. Although TACO derives a spatial map for memorization localization, its steering mechanism is not limited to local memorization and can also improve the mitigation of global memorization. Compared with baseline methods, TACO better suppresses memorized content while preserving prompt-specific semantic details. In the second example, it preserves the identity-related depiction of “John Doyle”, whereas the baseline simplifies the subject into text. In the fourth example, TACO better retains the structural details of the “foyer”, which are less clearly preserved by the baseline. These results highlight the broader applicability of TACO steering: it guides existing mitigation methods to suppress memorized content while better retaining relevant visual and semantic information, whether memorization affects a localized region or the image as a whole.

Figure 6. Qualitative comparison of existing memorization mitigation methods with and without TACO steering. Each pair compares the original mitigation baseline with its TACO-steered counterpart using the same generation seed, ensuring a controlled comparison.

Overall, the quantitative and qualitative results show that TACO provides a unified mechanism for both memorization localization and mitigation: it produces more precise spatial maps and uses them to improve the memorization-quality trade-off while retaining important prompt-specific details.

Ablation Study and Sensitivity Analysis

Ablation Study of TACO components. We conduct ablation experiments to evaluate the contributions of trajectory-aware aggregation and local refinement. Table 2 reports the results on SD v1.4 and SD v2.1 under both TV only and All settings. Removing trajectory aggregation and using a single observation step consistently reduces localization performance on TV only, indicating that combining curvature evidence across adjacent late steps provides more reliable spatial cues. Local refinement provides complementary improvements by sharpening the spatial localization of memorized regions. Removing this module decreases the performance, with a particularly noticeable mIoU drop on the All split of SD v2.1, from 0.923 to 0.905. This suggests that trajectory aggregation improves the reliability of memorization evidence, while local refinement helps translate that evidence into more accurate spatial boundaries.

ModelMethodTV onlyAll
mIoU(↑)ACC(↑)mIoU(↑)ACC(↑)
SD v1.4Ours0.9420.9640.9260.949
w/o Traj (only 48)0.9330.9590.9180.945
w/o Traj (only 49)0.9360.9600.9220.947
w/o Local Refine0.9390.9620.9230.948
SD v2.1Ours0.9570.9720.9230.977
w/o Traj (only 48)0.9490.9660.8940.970
w/o Traj (only 49)0.9510.9670.9260.973
w/o Local Refine0.9490.9660.9050.972

Table 2. Ablation study of TACO components.

Sensitivity to the Observation-Window Size ∣Tobs∣|\mathcal{T}_{\mathrm{obs}}|.

We study how the number of observation states affects TACO. Starting from the latest denoising states, we progressively expand the observation window toward earlier states. With sampling steps indexed from 0 to 49, the additional states are introduced in reverse order, such as 47, then 46, and so on. Thus, increasing ∣Tobs∣|\mathcal{T}_{\mathrm{obs}}| adds earlier states for curvature computation while retaining the previously selected late-stage observations. As shown in Figure 7, increasing the number of observed steps leads to an approximately linear increase in computation time, since curvature estimation must be performed at more points along the denoising trajectory. In contrast, the gains in mIoU and pixel-level accuracy diminish as the window expands. This trend is consistent with the observation that informative curvature changes are concentrated in the late denoising steps: adding earlier steps with less information incurs additional cost but contributes progressively less useful localization evidence. We therefore use ∣Tobs∣=2|\mathcal{T}_{\mathrm{obs}}| = 2, corresponding to steps {48,49}\{48, 49\}, as a practical balance between localization accuracy and computational efficiency.

Sensitivity analysis of the observation-window size

Figure 7. Sensitivity analysis of the observation-window size ∣Tobs∣|T_{\mathrm{obs}}|. The numbers annotated around each point indicate the corresponding window size.

Sensitivity to the Aggregation Weight.

{48,49}\{48, 49\}, we vary the weight wtstartw_{t_{\mathrm{start}}} assigned to t=48t = 48, while fixing wtend=1w_{t_{\mathrm{end}}} = 1 for the later state t=49t = 49. Setting wtstart=0w_{t_{\mathrm{start}}} = 0 retains only the response at the later state. As shown in Figure 8, increasing wtstartw_{t_{\mathrm{start}}} initially improves localization by incorporating the more spatially stable evidence from the earlier state. However, excessive weighting reduces the contribution of the stronger curvature contrast at the later state, leading to a decline in performance. The best results are obtained with wtstart=2w_{t_{\mathrm{start}}} = 2, highlighting the importance of balancing early-state spatial stability with late-state curvature sensitivity.

Figure 8. Sensitivity analysis of the aggregation weight wtstartw_{t_{\mathrm{start}}}.

Conclusion

In this paper, we present TACO, a trajectory-aware curvature observation framework for local memorization detection and mitigation in diffusion models. TACO aggregates curvature responses across neighboring late denoising steps and refines the resulting spatial evidence using latent-guided local refinement. The resulting continuous memorization map provides a unified interface for both tasks: it identifies memorized regions for localization and spatially guides existing inference-time interventions for mitigation. Extensive experiments demonstrate the effectiveness of TACO. Qualitative results further show that TACO helps preserve prompt-specific semantic details while directing mitigation toward memorization-related regions. Our findings highlight the potential of trajectory-aware spatial analysis as a unified approach to understanding and mitigating memorization in diffusion models.

Limitations and Future Works. Despite its effectiveness, estimating coordinate-wise curvature through vector-Jacobian products at multiple states introduces additional inference-time computation. Although TACO observes only a small number of late-stage states, this overhead may still limit its use in latency-sensitive applications. Future works could explore more efficient curvature estimators or probe-sharing strategies to reduce computational cost while retaining accurate localization and effective spatial steering.

Overview

This supplementary material provides background related works, methodological details, and more experimental details and results. Specifically:

  • Sec. B reviews prior works on diffusion-model memorization, its localization, and its mitigation.

  • Sec. C presents the pseudocode for TACO localization and steering, together with an analysis of its computational overhead.

  • Sec. D describes the experimental setup and reports additional quantitative and qualitative results, including quantitative Pareto-front comparisons, and more examples of memorization localization and TACO-steered mitigation.

Related Works

Memorization in Diffusion Models

Diffusion models generate images by reversing a progressive noising process [12, 29, 30]. In text-to-image (T2I) systems, latent diffusion combines this process with text conditioning [24], while classifier-free guidance (CFG) strengthens the conditional score during sampling [11]. Despite being trained to approximate a data distribution, these models can exhibit memorization, reproducing training images or generating highly similar outputs in response to certain prompts [3, 26, 32].

Memorization can differ in its spatial extent, leading to global and local memorization. Global memorization reproduces a training image at the image level, preserving its overall layout, object arrangement, and often its fine details. Local memorization reproduces only a spatially confined region, while the remaining parts of the generated image can vary across samples [5, 16]. The copied region may preserve near-verbatim details or retain a distinctive composition while changing colors, textures, and other surface attributes, corresponding to the template-like behavior reported in prior works [23, 32, 35]. Because local memorization can be hidden within an otherwise novel image, identifying its spatial extent is essential for detecting copied content and applying mitigation selectively.

Several factors make memorization easier to learn and retrieve. Repeated training examples receive greater effective weight during optimization, and specific captions can strengthen the association between an image and the words that retrieve it [3, 27]. Deduplication can reduce exact copying, but similar compositions and strong image-text associations may still support template memorization [9, 32]. These effects are also visible in the denoising network: memorized prompts often yield large conditional-unconditional score differences [33], while cross-attention can concentrate on a small number of prompt tokens associated with the copied content [23]. Memorization is further shaped by the sampling trajectory. [13] shows that CFG can create an attraction basin that steers a trajectory toward a memorized solution. Delaying or reversing guidance can help the trajectory leave this basin. [9] therefore treats the initial noise as an inference-time control variable and updates it using batch-wise finite differences or per-sample optimization.

A complementary line of works explains memorization through the geometry of the learned distribution. The Manifold Memorization Hypothesis [25] associates memorized samples with reduced local intrinsic dimensionality; [14] relates them to sharp conditional probability landscapes. [16] extends this perspective to spatial localization by estimating coordinate-wise curvature differences. These studies establish a progression from observing copied outputs to measuring conditioning-dependent, geometric, and spatial evidence.

Memorization Localization and Mitigation

Memorization detection methods typically ask whether a generated image contains memorized content. Retrieval-based methods compare the generated image with a training corpus using pixel- or feature-level similarity [3, 26]; self-supervised descriptors such as SSCD improve robustness to moderate transformations [21]. Retrieval provides direct evidence of exact or near-exact copying, but requires access to candidate training images and may miss local or template memorization when the copied region occupies only a small part of the output. Retrieval-free detectors instead use model responses, including score differences [33], local intrinsic dimensionality [25], and probability-landscape sharpness [14]. These signals are easier to obtain during sampling, but they generally produce image-level scores and do not indicate where the memorized content appears.

Localization addresses this complementary question by identifying the spatial regions associated with memorization. [23] associates memorization with concentrated attention on trigger or summary tokens, while Bright Ending [5] uses late-stage cross-attention to construct a spatial map. However, attention may also reflect ordinary semantic alignment and can vary across layers, tokens, and denoising states. [16] instead estimates reference-subtracted

Conclusion

coordinate-wise curvature, providing a geometric localization signal at a selected denoising state. TACO follows this coordinate-wise perspective and aggregates corresponding curvature responses across neighboring late-stage states, reducing transient peaks and producing more spatially consistent memorization maps.

Mitigation methods operate during training, after training, or at inference time. Training-time deduplication, data curation, and regularization reduce the chance of learning easily extractable copies [23, 27, 33]. Post-training editing and unlearning modify the model itself: concept-erasure and gradient-based methods suppress specified concepts [7, 8, 19], example-level unlearning targets selected training data [1], and representation-localization methods identify memorization subspaces, responsible neurons, or foreground-background mechanisms [4, 6, 10]. These approaches can produce persistent changes, but generally require target information and may alter unrelated capabilities. Inference-time methods preserve the released model and modify the prompt, guidance, initial noise, or attention pathway. Prompt perturbation and embedding optimization weaken associations between memorized images and their textual triggers [27, 33]. Guidance-based methods delay CFG, set it to zero, or apply opposite guidance to help trajectories escape memorization-attracting regions [13]. Han et al. [9] modify the initial noise through finite-difference or gradient-based updates. These methods control the sampling trajectory globally, so stronger suppression may also affect prompt-relevant regions. [35] introduces a more selective attention intervention. The detector finds token-wise attention spikes dynamically at each denoising step, attenuates the corresponding cross-attention blocks, and constructs an alternative conditional noise prediction.

TACO provides a spatial control interface that connects detection with inference-time mitigation. It aggregates coordinate-wise curvature evidence across neighboring late-stage states to form a continuous memorization map. The map is thresholded for localization and retained as a soft spatial weight for steering. In the experiments, TACO steers the initial-noise methods of [9] and the zero- or opposite-guidance strategies of [13]. Thus, TACO reuses existing mitigation objectives while directing their adjustable components toward regions supported by persistent spatial evidence.

Method Details

Algorithm Pseudocode

Algo. 1 summarizes the construction of the continuous TACO memorization map. Given a prompt, TACO follows the standard denoising trajectory and estimates prompt-dependent curvature only at the selected late-stage observation states. The resulting curvature responses are processed and spatially refined before being aggregated into a single memorization map.

Algorithm 1 TACOLocalization

Require: Text prompt pp, observation states Tobs\mathcal{T}_{\mathrm{obs}}

Ensure: Continuous memorization map MM

1: Sample initial latent x0x_0 and initialize C←∅\mathcal{C} \leftarrow\varnothing 2: for t=0,…,T−1t=0,\ldots,T-1 do 3: Evaluate ϵtu(xt)\epsilon_t^{u}(x_t) and ϵtc(xt,p)\epsilon_t^{c}(x_t,p) 4: if t∈Tobst \in\mathcal{T}_{\mathrm{obs}} then 5: rt←ϵtc(xt,p)−ϵtu(xt)r_t \leftarrow\epsilon_t^{c}(x_t,p)-\epsilon_t^{u}(x_t) 6: H^t←DiagonalCurvature⁡(xt,rt,σt)\widehat{H}_t \leftarrow\operatorname{DiagonalCurvature}(x_t,r_t,\sigma_t) Estimate diagonal curvature 7: Ht′(x,c)←ReLU⁡(−H^t(x,c))H_t^{\prime}(x,c) \leftarrow\operatorname{ReLU}(-\widehat{H}_t(x,c)) 8: Ct(x)←1C∑c=1CHt′(x,c)C_t(x) \leftarrow\frac{1}{C}\sum_{c=1}^{C} H_t^{\prime}(x,c) Aggregate negative curvature across channels 9: μtc,stc←SpatialStatistics⁡(Ct)\mu_t^{c},s_t^{c} \leftarrow\operatorname{SpatialStatistics}(C_t) 10: C~t(x)←Ct(x)sigmoid⁡(βCt(x)−μtcstc+ϵ)\widetilde{C}_t(x) \leftarrow C_t(x)\operatorname{sigmoid}\left(\frac{\beta C_t(x)-\mu_t^{c}}{s_t^{c}+\epsilon}\right) Emphasize salient responses 11: C~t←LatentGuidedRefinement⁡(Ct,xt)\widetilde{C}_t \leftarrow\operatorname{LatentGuidedRefinement}(C_t,x_t) 12: C←C∪{(t,C~t)}\mathcal{C} \leftarrow\mathcal{C}\cup\{(t,\widetilde{C}_t)\} 13: end if 14: ϵtg←ϵtu+s(ϵtc−ϵtu)\epsilon_t^{g} \leftarrow\epsilon_t^{u}+s(\epsilon_t^{c}-\epsilon_t^{u}) 15: xt+1←Dt(xt,ϵtg)x_{t+1}\leftarrow\mathcal{D}_t(x_t,\epsilon_t^{g}) 16: end for 17: Sort Tobs={t1,…,tn}\mathcal{T}_{\mathrm{obs}}=\{t_1,\ldots,t_n\} in ascending order 18: for i=1,…,ni=1,\ldots,n do 19: wi←2−i−1n−1w_i\leftarrow2-\frac{i-1}{n-1} 20: end for 21: M←∑i=1nwiC~ti∑i=1nwiM\leftarrow\frac{\sum_{i=1}^{n}w_i\widetilde{C}_{t_i}}{\sum_{i=1}^{n}w_i} 22: return MM

Algorithm 2 summarizes the TACO steering mechanism for memorization mitigation. Given a continuous memorization map and an intervention produced by an arbitrary baseline method, when the baseline objective gradient is available, TACO preserves the objective-aligned intervention component and spatially redistributes only its orthogonal component according to the memorization map. Otherwise, TACO combines localized attenuation of classifier-free guidance with a transverse correction orthogonal to the prompt-dependent score difference. In both cases, TACO concentrates the intervention on memorization-related regions while preserving the baseline method’s original optimization role.

Computational Overhead

Let TT denote the number of sampling steps, n=∣Tobs∣n=|\mathcal{T}_{\mathrm{obs}}| the number of observed states, JJ the number of curvature probes, and D=C×H×WD=C\times H\times W the latent dimensionality. Let FF and VV denote the costs of one U-Net forward evaluation and one vector-Jacobian product, respectively. Standard Algorithm 2 TACOSTEERING

Require: Text prompt pp, baseline intervention Δtbase\Delta^{\mathrm{base}}_{t}, current latent xtx_{t}

Ensure: TACO-steered intervention ΔtTACO\Delta^{\mathrm{TACO}}_{t}

1: M←TACOLOCALIZATION⁡(p)M \leftarrow\operatorname{TACOLOCALIZATION}(p) 2: M~t←PREPROCESSMAP⁡(M,xt)\widetilde{M}_{t} \leftarrow\operatorname{PREPROCESSMAP}(M, x_{t}) ▷ Preprocessing map 3: if objective gradient gtg_{t} is available then 4: Δt∥←⟨Δtbase,gt⟩∥gt∥22+ϵgt,Δt⊥←Δtbase−Δt∥\Delta^{\parallel}_{t} \leftarrow\frac{\langle\Delta^{\mathrm{base}}_{t},g_{t}\rangle}{\lVert g_{t}\rVert_{2}^{2}+\epsilon}g_{t}, \quad\Delta^{\perp}_{t} \leftarrow\Delta^{\mathrm{base}}_{t}-\Delta^{\parallel}_{t} 5: Δ^t⊥←M~t⊙Δt⊥\widehat{\Delta}^{\perp}_{t} \leftarrow\widetilde{M}_{t}\odot\Delta^{\perp}_{t} 6: if ∥Δ^t⊥∥2>ϵ\lVert\widehat{\Delta}^{\perp}_{t}\rVert_{2}>\epsilon then 7: Δt⊥,TACO←∥Δt⊥∥2∥Δ^t⊥∥2+ϵΔ^t⊥\Delta^{\perp,\mathrm{TACO}}_{t}\leftarrow\frac{\lVert\Delta^{\perp}_{t}\rVert_{2}}{\lVert\widehat{\Delta}^{\perp}_{t}\rVert_{2}+\epsilon}\widehat{\Delta}^{\perp}_{t} 8: ΔtTACO←Δt∥+Δt⊥,TACO\Delta^{\mathrm{TACO}}_{t}\leftarrow\Delta^{\parallel}_{t}+\Delta^{\perp,\mathrm{TACO}}_{t} 9: else 10: ΔtTACO←Δtbase\Delta^{\mathrm{TACO}}_{t}\leftarrow\Delta^{\mathrm{base}}_{t} 11: end if 12: else 13: rt←ϵtc(xt,p)−ϵtu(xt)r_{t}\leftarrow\epsilon^{c}_{t}(x_{t},p)-\epsilon^{u}_{t}(x_{t}) 14: At←MaskedRMS⁡(rt,M~t)A_{t}\leftarrow\operatorname{MaskedRMS}(r_{t},\widetilde{M}_{t}) ▷ Measure the map-weighted guidance strength 15: qt←∇xtAtq_{t}\leftarrow\nabla_{x_{t}}A_{t} 16: qt∥←⟨qt,rt⟩∥rt∥22+ϵrt,qt⊥←qt−qt∥q^{\parallel}_{t}\leftarrow\frac{\langle q_{t},r_{t}\rangle}{\lVert r_{t}\rVert_{2}^{2}+\epsilon}r_{t}, \quad q^{\perp}_{t}\leftarrow q_{t}-q^{\parallel}_{t} 17: Δtlocal←−λcfgM~t⊙rt\Delta^{\mathrm{local}}_{t}\leftarrow-\lambda_{\mathrm{cfg}}\widetilde{M}_{t}\odot r_{t} ▷ Locally attenuate classifier-free guidance 18: Δttrans←σtAtqt⊥RMS⁡(qt)+ϵ\Delta^{\mathrm{trans}}_{t}\leftarrow\sigma_{t}A_{t}\frac{q^{\perp}_{t}}{\operatorname{RMS}(q_{t})+\epsilon} ▷ Apply the normalized transverse correction 19: ΔtTACO←Δtbase+Δtlocal+Δttrans\Delta^{\mathrm{TACO}}_{t}\leftarrow\Delta^{\mathrm{base}}_{t}+\Delta^{\mathrm{local}}_{t}+\Delta^{\mathrm{trans}}_{t} 20: end if 21: return ΔtTACO\Delta^{\mathrm{TACO}}_{t}

dard sampling requires O(TF)O(TF) computation. TACO localization additionally performs forward evaluations at the observed states, probe-based curvature estimation, and latent-guided refinement over k×kk \times k neighborhoods, yielding an additional cost of

O(nF+nJV+nDk2).(9)O(nF+nJV+nDk^{2}). \tag*{(9)}

In our implementation, n=2n=2, J=16J=16, and k=3k=3. Curvature estimation therefore requires only 32 structured vector-Jacobian products, while the remaining map-processing operations are inexpensive relative to U-Net evaluation. The probes are evaluated sequentially, so the accumulated curvature map requires O(D)O(D) additional storage, apart from the autograd graph used by the current observation state.

TACO steering is substantially lighter. Let D=C×H×WD=C\times H\times W denote the latent dimensionality and let LL denote the number of intervention steps. Apart from the map construction process, when the objective gradient is available, TACO

Table 3. Average generation time per image in seconds. Values in parentheses denote the additional runtime introduced by TACO calculation steering.

MethodSD v1.4SD v2.0
Base+OursBase+Ours
No Mitigation0.97-0.96-
Init Batch1.031.27 (+0.24)0.971.20 (+0.23)
Init Sample2.042.40 (+0.36)1.792.05 (+0.26)
Basin Zero0.951.26 (+0.31)0.901.23 (+0.33)
Basin Neg0.951.26 (+0.31)0.901.20 (+0.30)

Table 3.

steering consists of element-wise spatial reweighting, inner products, and norm rescaling. It therefore incurs O(D)O(D) additional computation at each intervention step, or O(LD)O(LD) over the full sampling process. When the objective gradient is unavailable, the transverse correction additionally requires one gradient evaluation of the map-weighted attraction objective at each applicable intervention step. Apart from this gradient computation, all steering operations remain linear in the latent dimensionality. Thus, TACO can be integrated with different mitigation objectives without altering their optimization procedure or introducing substantial additional memory overhead.

Tab. 3 reports the average wall-clock generation time per image. In our mitigation evaluation, each prompt produces 10 samples using seeds 0 through 9. The TACO memorization map is computed only once from the generation with seed 0 and is subsequently reused for the remaining samples of the same prompt. Thus, the curvature-based localization cost is shared across all ten generations, whereas the reported additional runtime mainly reflects the lightweight steering operations.

Across the four mitigation baselines, TACO introduces an additional 0.24-0.36 seconds per image on SD v1.4 and 0.23-0.33 seconds on SD v2.0. The overhead remains consistent despite the different computational profiles of the baseline methods, indicating that TACO steering acts as an efficient plug-and-play module. The prompt-level reuse of the TACO map further reduces the effective cost of spatial localization during multi-sample mitigation.

Experimental Details and Additional Results

Experimental Setup

Models and Runtime Environment.

Following the corresponding benchmark settings, we evaluate memorization localization using SD v1.4 (CompVis/stable-diffusion-v1-4) and SD v2.1 (Manojb/stable-diffusion-2-1-base). For memorization mitigation, we use SD v1.4 and SD v2.0 (Manojb/stable-diffusion-2-base). In both localization and mitigation settings, we use a DDIM scheduler with 50 inference steps and a CFG guidance scale of 7.5. Sampling steps are indexed according to their execution order from 0 to 49. All experiments are implemented in PyTorch 2.4.0 [20] using the Hugging Face Diffusers library (version 0.36.0) [31] and run under Ubuntu 22.04.5 LTS on a server equipped with one single NVIDIA GeForce RTX 4090 GPU with 24GB VRAM.

TACO for Memorization Localization. We evaluate memorization localization on the benchmark provided by [16]. Each benchmark instance consists of a prompt–seed pair and its corresponding ground-truth binary mask, where a value of one indicates a memorized region and zero indicates a non-memorized region. For each prompt, we evaluate at most four generation seeds provided by the benchmark. The benchmark is evaluated under two settings: TV only, which contains samples with localized memorized regions, and All, which additionally includes non-memorized samples (For non-memorized samples, The generation seed is fixed at 0). The latter setting evaluates whether a method can distinguish memorized content from ordinary generated regions.

We use mean Intersection over Union (mIoU) and pixel-level accuracy (Acc) to quantify memorization localization. Given a predicted continuous memorization map MiM_i and its ground-truth binary mask YiY_i for sample ii, we obtain a binary prediction mask using a threshold τ\tau:

Y^i,τ(x)=I[Mi(x)>τ],(10)\widehat{Y}_{i,\tau}(x) = \mathbb{I}[M_i(x) > \tau], \tag*{(10)}

where Yi(x)∈{0,1}Y_i(x) \in\{0, 1\} indicates whether spatial location xx belongs to a memorized region. For each sample, the Intersection over Union (IoU) is defined as IoU⁡i=∣Y^i∩Yi∣∣Y^i∪Yi∣\operatorname{IoU}_i = \frac{|\widehat{Y}_i \cap Y_i|}{|\widehat{Y}_i \cup Y_i|}. The mean IoU averages the per-sample IoU over the NN evaluation samples: mIoU⁡=1N∑i=1NIoU⁡i\operatorname{mIoU} = \frac{1}{N}\sum_{i=1}^{N}\operatorname{IoU}_i. Pixel-level accuracy measures the fraction of correctly classified pixels across all samples as Acc⁡=∑i=1N∑xI[Y^i(x)=Yi(x)]∑i=1N∑x1\operatorname{Acc} = \frac{\sum_{i=1}^{N}\sum_x \mathbb{I}[\widehat{Y}_i(x)=Y_i(x)]}{\sum_{i=1}^{N}\sum_x 1}.

For each metric, the threshold τ\tau is selected independently to optimize the corresponding metric on its evaluation set. We report both mIoU and Acc to assess complementary aspects of localization quality: mIoU emphasizes spatial overlap, whereas Acc measures overall pixel-wise classification correctness.

TACO steering for Memorization Mitigation. We use the benchmark provided by [32], which contains text prompts associated with both local and global memorization. For each prompt, we generate 10 samples using seeds 0 through 9, and compare each mitigated sample with the reference image. To avoid repeated computation, we construct the TACO map once from the sampling trajectory generated with seed 0. The same map is then reused to steer mitigation for all ten seeds of that prompt.

We integrate TACO steering into several existing mitigation methods to evaluate its effectiveness across different intervention strategies. Unless otherwise specified, we follow the default configurations of each baseline.

To characterize the trade-off between memorization suppression and generation quality, we vary the mitigation strength of each baseline and evaluate each configuration both with and without TACO steering:

  • For Init Batch, we vary the intervention step over {0,1,2,3,4,5,6}\{0,1,2,3,4,5,6\}, with 4 as the default; earlier interventions generally favor quality preservation.

  • For Init Sample, we vary the learning rate for initial-noise optimization and the early-stopping target loss, denoted as the (lr,target loss)(\text{lr}, \text{target loss}) pair. We evaluate {(0.006,1.2),(0.006,0.9),(0.008,0.9),(0.010,0.9),(0.012,0.9),(0.014,0.9),(0.014,0.6)}\{(0.006,1.2),(0.006,0.9),(0.008,0.9),(0.010,0.9),(0.012,0.9),(0.014,0.9),(0.014,0.6)\}, with (0.01,0.9)(0.01,0.9) as the default. Lower learning rates and higher target losses generally favor quality preservation.

  • For Basin Zero and Basin Neg, we vary the intervention duration in the diffusion timestep scale over {50,100,150,200,250,300,350}\{50,100,150,200,250,300,350\} and {50,75,85,100,125,150,175,200}\{50,75,85,100,125,150,175,200\}, respectively, with 100 as the default for both methods. An intervention duration of dd corresponds to intervening over the initial dd diffusion timesteps (e.g., from timestep 1000 to 1000−d1000-d). Shorter durations generally better preserve image quality.

We compare the resulting Pareto frontiers in the SSCD-CLIP and SSCD-ImageReward spaces to assess whether TACO steering improves memorization suppression while preserving semantic alignment and generation quality. For SSCD, we use the pretrained checkpoint sscd_disc_large.torchscript.pt released by [21], which extracts image descriptors for measuring visual similarity between each mitigated image and its corresponding unmitigated reference. Lower SSCD similarity indicates stronger suppression of memorized content. For semantic alignment, we use the ViT-B-32 model pretrained on laion2b_s34b_b79k to measure image-text similarity between the generated image and its input prompt. For perceptual preference quality, we use the pretrained ImageReward-v1.0 model, which predicts human preference for text-to-image generations. Higher CLIP and ImageReward scores indicate better semantic alignment and perceptual preference, respectively.

To quantitatively compare the resulting Pareto frontiers, we report the normalized Pareto hypervolume difference (NPHV) based on the hypervolume indicator [36]:

NPHV⁡=HVours−HVbaseHVours+HVbase,(11)\operatorname{NPHV} = \frac{HV_{\mathrm{ours}} - HV_{\mathrm{base}}}{HV_{\mathrm{ours}} + HV_{\mathrm{base}}}, \tag*{(11)}

where HVoursHV_{\mathrm{ours}} and HVbaseHV_{\mathrm{base}} denote the hypervolumes of the TACO-steered mitigation method and its corresponding original baseline, respectively. For each objective pair, we jointly apply min-max normalization to both methods using shared minimum and maximum values; the objective to be minimized (SSCD) are reversed as 1−x1 - x, while objectives to be maximized (CLIP and ImageReward) retain their normalized values. The hypervolumes are computed using a common reference point (−0.01,−0.01)(-0.01, -0.01), slightly below the normalized objective space, and measure the area covered by the rectangles between this point and the non-dominated solutions. A positive NPHV indicates that the proposed method achieves a larger hypervolume than the baseline, while a negative value indicates the opposite.

We additionally report the Pareto dominance advantage (PDA), which measures the fraction of the proposed method’s operating points that dominate at least one baseline operating point:

PDA=∑p∈PoursI[∃q∈Pbase:p≻q]∣Pours∣,(12)\mathrm{PDA} = \frac{\sum_{p \in\mathcal{P}_{\mathrm{ours}}} \mathbb{I}[\exists q \in\mathcal{P}_{\mathrm{base}} : p \succ q]}{|\mathcal{P}_{\mathrm{ours}}|}, \tag*{(12)}

where Pours\mathcal{P}_{\mathrm{ours}} and Pbase\mathcal{P}_{\mathrm{base}} denote the respective sets of operating points, and p≻qp \succ q indicates that pp achieves no worse SSCD and no worse quality score than qq, with at least one strict improvement. A higher PDA indicates that a larger proportion of the proposed method’s operating points dominate at least one baseline operating point. The corresponding results are shown in Table 4.

ModelMethod Pair (ours − base)SSCD-CLIP NPHV(↑)SSCD-CLIP PDA(↑)SSCD-ImageReward NPHV(↑)SSCD-ImageReward PDA(↑)
SD v1.4Init Batch0.0610.7140.1450.714
SD v1.4Init Sample0.1361.0000.1811.000
SD v1.4Basin Zero0.0650.2850.0910.714
SD v1.4Basin Neg0.0420.8570.0610.857
SD v2.0Init Batch0.2741.0000.3071.000
SD v2.0Init Sample0.2301.0000.1351.000
SD v2.0Basin Zero0.1030.5710.1170.571
SD v2.0Basin Neg0.0630.2850.0820.285

Table 4. Quantitative evaluation of Pareto fronts in memorization mitigation.

Additional Results

Table 4 summarizes the Pareto-front comparison between each baseline and its TACO-steered counterpart. All eight comparisons yield positive NPHV values in both evaluation spaces, indicating that TACO consistently enlarges the trade-off region between memorization suppression and generation quality. For SD v1.4, the largest improvements are obtained with Init Sample, whose NPHV reaches 0.136 in SSCD-CLIP and 0.181 in SSCD-ImageReward. Its PDA is 1.000 in both spaces, meaning that every TACO-steered operating point dominates at least one point of the corresponding baseline frontier. The other three baselines also benefit from TACO. These results show that the spatial map remains useful across both initial-noise and guidance-based interventions.

The improvements are even larger for SD v2.0. TACO-steered Init Batch obtains the highest NPHV, with gains of 0.274 in SSCD-CLIP and 0.307 in SSCD-ImageReward. Init Sample also achieves substantial improvements of 0.230 and 0.135, respectively, with PDA equal to 1.000 in both spaces. The two guidance-based methods remain consistently improved. Thus, the benefit of TACO is observed across different mitigation mechanisms and models, rather than being tied to a particular baseline.

Overall, the uniformly positive NPHV and PDA values demonstrate that TACO improves the suppression-quality trade-off across all evaluated settings. The results support the effectiveness of using the trajectory-aggregated spatial memorization map to steer existing mitigation objectives toward the regions most strongly associated with memorized content.

Table 5–Table 8 provide the detailed SSCD, CLIP, and ImageReward scores for the four groups respectively. These tables report the operating points used to construct the Pareto frontiers and provide the underlying quantitative results for the comparisons summarized in Table 4.

MethodParamSD v1.4SD v2.0
SSCD(↓\downarrow)CLIP(↑\uparrow)ImageReward(↑\uparrow)SSCD(↓\downarrow)CLIP(↑\uparrow)ImageReward(↑\uparrow)
No mitigation-0.51060.29990.45480.31960.30390.1406
Init Batch00.38070.2876−0.14190.28990.2982−0.0033
Init Batch10.34410.2812−0.20500.28090.2954−0.0528
Init Batch20.30610.2743−0.28090.26880.2919−0.1260
Init Batch30.27000.2670−0.36320.25340.2865−0.2109
Init Batch40.23760.2604−0.43770.23440.2824−0.2839
Init Batch50.20500.2522−0.53720.21530.2771−0.3662
Init Batch60.17770.2452−0.62420.19350.2714−0.4606
Init Batch+Ours00.34730.2821−0.15920.28020.29910.0363
Init Batch+Ours10.30520.2754−0.22160.26810.2970−0.0144
Init Batch+Ours20.26760.2700−0.28200.25380.2944−0.0584
Init Batch+Ours30.23400.2633−0.36040.23600.2903−0.1280
Init Batch+Ours40.20500.2568−0.43460.21740.2862−0.1905
Init Batch+Ours50.17680.2499−0.51850.19880.2823−0.2666
Init Batch+Ours60.15530.2436−0.59740.18030.2768−0.3613

Table 5. Detailed statistics of Pareto fronts comparison between the original Init Batch and TACO-steered Init Batch.

MethodParamSD v1.4SD v2.0
SSCD(↓\downarrow)CLIP(↑\uparrow)ImageReward(↑\uparrow)SSCD(↓\downarrow)CLIP(↑\uparrow)ImageReward(↑\uparrow)
No mitigation-0.51060.29990.45480.31960.30390.1406
Basin Neg500.45500.2955−0.04800.30430.30220.0825
750.41100.2905−0.10210.29210.29970.0320
850.34790.2815−0.19990.27210.2946−0.0581
1000.28010.2687−0.33890.23700.2866−0.2039
1250.20850.2534−0.51390.19360.2751−0.3836
1500.15200.2360−0.71990.14820.2610−0.6138
1750.10980.2191−0.91890.11360.2466−0.8275
Basin Neg+Ours500.40950.2918−0.06730.28860.30040.0595
750.33800.2830−0.14790.26610.29780.0153
850.26800.2712−0.27350.23970.2914−0.0983
1000.19640.2562−0.44700.19620.2806−0.2621
1250.14290.2392−0.65800.15530.2683−0.4703
1500.10190.2224−0.87160.11720.2534−0.7219
1750.07710.2076−1.08180.08930.2391−0.9440

Table 8. Detailed statistics of Pareto fronts comparison between the original Basin Neg and TACO-steered Basin Neg.

Additional Visualizations

Figure 9 compares the memorization maps produced by TACO and other localization methods. TACO generates spatially coherent responses with clearer boundaries around the memorized regions, while reducing activations in surrounding non-memorized areas. Figure 10 shows the final generated images obtained with and without TACO steering for several existing mitigation methods. The TACO-steered results reduce the copied content highlighted by the map while preserving the surrounding image structure. Together, these examples demonstrate that TACO provides both interpretable local memorization maps and an effective spatial control signal for guiding existing mitigation methods.

Additional qualitative comparison of memorization localization maps produced by different methods

Figure 9. Additional qualitative comparison of memorization localization maps produced by different methods.

Figure 10. Additional qualitative comparison of existing memorization mitigation methods with and without TACO steering. Each pair compares the original mitigation baseline with its TACO-steered counterpart using the same generation seed.

MethodParamSD v1.4SD v2.0
SSCD(↓\downarrow)CLIP(↑\uparrow)ImageReward(↑\uparrow)SSCD(↓\downarrow)CLIP(↑\uparrow)ImageReward(↑\uparrow)
No mitigation-0.51060.2999−0.00600.31960.30390.1406
Init Sample(0.006, 1.2)0.29030.2807−0.18290.25930.29810.0628
Init Sample(0.006, 0.9)0.22110.2697−0.30810.22830.2887−0.0763
Init Sample(0.008, 0.9)0.20850.2673−0.32530.21720.2867−0.1111
Init Sample(0.010, 0.9)0.19820.2653−0.35100.20910.2859−0.1181
Init Sample(0.012, 0.9)0.18450.2618−0.38940.19710.2844−0.1447
Init Sample(0.014, 0.9)0.17050.2592−0.42370.18620.2828−0.1870
Init Sample(0.014, 0.6)0.09870.2328−0.78710.12750.2634−0.5711
Init Sample+Ours(0.006, 1.2)0.28620.2822−0.14710.25660.30010.0804
Init Sample+Ours(0.006, 0.9)0.20630.2719−0.24360.21030.2936−0.0522
Init Sample+Ours(0.008, 0.9)0.18930.2691−0.27180.20120.2923−0.0843
Init Sample+Ours(0.010, 0.9)0.17800.2680−0.27640.19030.2905−0.0998
Init Sample+Ours(0.012, 0.9)0.16600.2650−0.30450.17970.2898−0.1253
Init Sample+Ours(0.014, 0.9)0.15790.2635−0.33120.17130.2884−0.1426
Init Sample+Ours(0.014, 0.6)0.09660.2403−0.64220.11900.2674−0.5581

Table 6. Detailed statistics of Pareto fronts comparison between the original Init Sample and TACO-steered Init Sample.

MethodParamSD v1.4SD v2.0
SSCD(↓\downarrow)CLIP(↑\uparrow)ImageReward(↑\uparrow)SSCD(↓\downarrow)CLIP(↑\uparrow)ImageReward(↑\uparrow)
No mitigation-0.51060.29990.45480.31960.30390.1406
Basin Zero500.48710.2984−0.02590.31320.30270.1097
1000.44100.2926−0.08010.30090.30030.0411
1500.39090.2856−0.16540.28470.2967−0.0369
2000.29750.2685−0.36870.24310.2868−0.2177
2500.23520.2545−0.53050.19910.2766−0.3889
3000.15220.2294−0.83680.13420.2570−0.7036
3500.11060.2121−1.07050.09810.2408−0.9441
Basin Zero+Ours500.45020.2965−0.02420.29880.30200.0970
1000.32920.2821−0.16280.25920.29820.0204
1500.24960.2681−0.30990.22360.2912−0.1073
2000.14320.2412−0.62170.15600.2749−0.3902
2500.09720.2243−0.84690.11180.2607−0.6316
3000.06180.1990−1.18690.07090.2383−0.9959
3500.04940.1830−1.41660.05920.2240−1.2163

Table 7. Detailed statistics of Pareto fronts comparison between the original Basin Zero and TACO-steered Basin Zero.

References

  1. [1]Silas Alberti, Kenan Hasanaliyev, Manav Shah, and Stefano Ermon. Data unlearning in diffusion models. In International Conference on Learning Representations, pages 3084–3100, 2025. 1, 2arxiv.org/abs/2503.01034
  2. [2]Costas Bekas, Efrosyni Kokiopoulou, and Yousef Saad. An estimator for the diagonal of a matrix. Applied numerical mathematics, 57(11-12):1214–1229, 2007. 3DOI
  3. [3]Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253–5270, 2023. 1arxiv.org/abs/2301.13188
  4. [4]Ruchika Chavhan, Ondrej Bohdal, Yongshuo Zong, Da Li, and Timothy Hospedales. Memorized images in diffusion models share a subspace that can be located and deleted. arXiv preprint arXiv:2406.18566, 2024. 1, 2arxiv.org/abs/2406.18566
  5. [5]Chen Chen, Daochang Liu, Mubarak Shah, and Chang Xu. Exploring local memorization in diffusion models via bright ending attention. In The Thirteenth International Conference on Learning Representations, 2025. 1, 5arxiv.org/abs/2410.21665
  6. [6]Jimmy Z Di, Yiwei Lu, Yaoliang Yu, Gautam Kamath, Adam Dziędzic, and Franziska Boenisch. Demystifying foreground-background memorization in diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 20763–20771, 2026. 1, 2
  7. [7]Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Sulan: Empowering machine un-learning via gradient-based weight saliency in both image classification and generation. In International Conference on Learning Representations, pages 53643–53673, 2024. 2
  8. [8]Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426–2436, 2023. 1, 2arxiv.org/abs/2303.07345
  9. [9]Hyeonggeun Han, Sehwan Kim, Hyungjun Joo, Sangwoo Hong, and Jungwoo Lee. Adjusting initial noise to mitigate memorization in text-to-image diffusion models. In Advances in Neural Information Processing Systems, pages 684–711. Curran Associates, Inc., 2025. 2, 5, 1arxiv.org/abs/2510.08625
  10. [10]Dominik Hintersdorf, Lukas Struppek, Kristian Kersting, Adam Dziedzic, and Franziska Boenisch. Finding nemo: Localizing neurons responsible for memorization in diffusion models. Advances in Neural Information Processing Systems, 37:88236–88278, 2024. 1, 2arxiv.org/abs/2406.02366
  11. [11]Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 1arxiv.org/abs/2207.12598
  12. [12]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020. 1arxiv.org/abs/2006.11239
  13. [13]Anubhav Jain, Yuya Kobayashi, Takashi Shibuya, Yuhta Takida, Nasir Memon, Julian Tegelius, and Yuki Mitsufuji. Classifier-free guidance inside the attraction basin may cause memorization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871–12879, 2025. 2, 5, 1arxiv.org/abs/2411.16738
  14. [14]Dongjae Jeon, Dueun Kim, and Albert No. Understanding and mitigating memorization in generative models via sharpness of probability landscapes. arXiv preprint arXiv:2412.04140, 2024. 1arxiv.org/abs/2412.04140
  15. [15]Hyungjun Joo, Sehwan Kim, Hyeonggeun Han, Sangwoo Hong, and Jungwoo Lee. Rethinking memorization mitigation in diffusion models: Reinforcing text conditioning. arXiv preprint arXiv:2610.01723, 2026. 2arxiv.org/abs/2610.01723
  16. [16]Gwangho Kim and Sungyoon Lee. Localizing memorized regions in diffusion models via coordinate-wise curvature differences. arXiv preprint arXiv:2605.26756, 2026. 2, 3, 5, 6, 1, 4arxiv.org/abs/2605.26756
  17. [17]Antoni Kowalczuk, Dominik Hintersdorf, Lukas Struppek, Kristian Kersting, Adam Dziedzic, and Franziska Boenisch. Finding dori: Memorization in text-to-image diffusion models is less local than assumed. arXiv e-prints, pages arXiv–2507, 2025. 1
  18. [18]Kaisa Miettinen et al. Nonlinear multiobjective optimization. Kluwer Academic Publishers Boston, 1999. 2
  19. [19]Yong-Hyun Park, Sangdoo Yun, Jin-Hwa Kim, Junho Kim, Geonhui Jang, Yonghyun Jeong, Junghyo Jo, and Gayoung Lee. Direct unlearning optimization for robust and safe text-to-image models. Advances in Neural Information Processing Systems, 37:80244–80267, 2024. 1, 2arxiv.org/abs/2407.21035
  20. [20]Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024–8035. Curran Associates, Inc., 2019. 4arxiv.org/abs/1912.01703
  21. [21]Ed Pizzi, Sreya Dutta Roy, Sugosh Nagarava Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532–14542, 2022. 1, 5, 4arxiv.org/abs/2202.10261
  22. [22]Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021. 5arxiv.org/abs/2103.00020
  23. [23]Jie Ren, Yaxin Li, Shengli Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, pages 340–356. Springer, 2024. 1, 2arxiv.org/abs/2403.11052
  24. [24]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 1, 2, 5arxiv.org/abs/2112.10752
  25. [25]Brendan Ross, Hamidreza Kamkari, Tongzi Wu, Rasa Hosseinzadeh, Zhaoyan Liu, George Stein, Jesse Cresswell, and Gabriel Loaiza-Ganem. A geometric framework for un derstanding memorization in generative models. In International Conference on Learning Representations, pages 34773–34805, 2025. 1
  26. [26]Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048–6058, 2023. 1
  27. [27]Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 1arxiv.org/abs/2305.20086
  28. [28]Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5arxiv.org/abs/2010.02502
  29. [29]Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1arxiv.org/abs/2010.02502
  30. [30]Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1
  31. [31]Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022. 4
  32. [32]Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 1, 5, 4arxiv.org/abs/2305.08694
  33. [33]Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2arxiv.org/abs/2407.21720
  34. [34]Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagerward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903–15935, 2023. 5arxiv.org/abs/2304.05977
  35. [35]Kairan Zhao, Eleni Triantafillou, and Peter Triantafillou. You don’t need all that attention: Surgical memorization mitigation in text-to-image diffusion models. arXiv preprint arXiv:2603.00133, 2026. 2, 1arxiv.org/abs/2603.00133
  36. [36]Eckart Zitzler, Dimo Brockhoff, and Lothar Thiele. The hypervolume indicator revisited: On the design of pareto-compliant indicators via weighted integration. In International conference on evolutionary multi-criterion optimization, pages 862–876. Springer, 2007. 4DOI

Paper details

vlnc:W_c9t8nn6j
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents